Your firm's own model

Every law firm using AI today is renting its intelligence. The model belongs to someone else, runs somewhere else, changes on someone else's schedule, and knows nothing about your firm that you didn't send it in the last five minutes. For most of the market, that's simply the deal — "model choice," where it exists at all, means choosing which company's cloud your privileged work is sent to.

There's a longer spectrum than that, and it's worth understanding even if your firm never walks to the end of it — because where a platform sits on this spectrum tells you what it believes about who should own the intelligence layer of a law practice.

The model freedom spectrum

There are three positions a firm can occupy, and they aren't exclusive — a firm can run all three at once, for different work.

• Frontier models, on your keys. The strongest general-purpose models, accessed through your firm's own API keys — you pay the provider directly, on terms you negotiated, with credentials stored encrypted in your deployment. Maximum capability; the tradeoff is that inference runs on someone else's infrastructure.
• Local open-source models, on your hardware. Capable open models running inside the firm's own deployment. Nothing leaves; token cost drops to zero at the margin. The tradeoff is that you're operating hardware and accepting open-model capability rather than frontier capability.
• The firm's own model. A local open-source model, fine-tuned on the firm's own curated work product, on the firm's own infrastructure — a model that is not just hosted by the firm but shaped by it, and owned by it.

EdgeLex supports all three, governed identically: a three-layer, default-deny model policy decides what may run at all, and a routing layer keeps underpowered models away from destructive actions regardless of which class they belong to. The third position — the firm's own model — is what the Lex Training Center exists for, and it's the one that deserves the most careful explanation, because it's the easiest to oversell. So let's be precise about what it is, how it works, and what it is not.

Why a firm would want its own model

Not for raw intelligence — we'll get to that. The honest reasons are narrower and more interesting.

Voice and house style. Every firm has one: how a demand letter opens, how a client update is structured, which hedges the senior partners strike and which they insist on. A general model approximates "a lawyer"; a tuned model can be taught, from curated examples of the firm's own accepted work, to sound like your firm — so drafts start closer to done and partners spend less time rewriting competent-but-foreign prose into something they'd sign.

Your practice's patterns. The recurring motions, the standard client situations, the way your firm handles the matters it handles most — a curated dataset built from that work teaches a model the shape of your practice specifically, rather than the average of everyone's.

Privacy, completed. Running local models already keeps inference inside your walls. A firm-tuned model closes the remaining loop: the improvement process itself — the data, the training, the resulting weights — also never leaves the deployment. The firm's accumulated judgment stays the firm's asset, rather than dissolving into a vendor's product.

And there's a structural point underneath all three: a firm that can run its own model can never be wholly captive to any model vendor's pricing, terms, or roadmap. The option itself is leverage, even on days you don't exercise it.

It's worth saying what the underlying data is, because it's the part firms rightly guard most closely. The raw material for a firm model is the firm's own work product — the drafts partners accepted, the answers lawyers corrected, the examples someone deliberately wrote down as "this is how we do it." That corpus is arguably the most valuable intangible a firm owns, which is exactly why the default everywhere in EdgeLex is that it is never training material. Contracts, matters, and emails are never used to train models unless the firm chooses to build a dataset from them — and when it does, the choosing happens example by example, through the pipeline described next.

How the Training Center actually works

The path from daily work to a firm model runs through four deliberate stages — and the deliberateness is the design. Nothing in this pipeline happens to your firm; every stage is something your firm does, with review gates between.

Feedback is signal, not training data

It starts where lawyers already work. The Training Center captures feedback from real use — thumbs up and down, inline corrections, rephrase detection, abandoned topics, natural follow-ups — building a living record of where Lex helped and where it fell short. But feedback is signal-only. Nothing auto-trains. A thumbs-down never silently rewires the model. Feedback tells the humans curating the dataset where to look; it never becomes training data on its own.

Nothing auto-trains. A thumbs-down never silently rewires the model — feedback tells your lawyers where to look, and lawyers decide what the model learns.

Lawyers curate; gates fail closed

Datasets grow only two ways: lawyer-recreated manual examples, or completed work traces a person explicitly selects. Candidates then pass through the Clean-Baseline Training Factory, which produces named datasets with full lineage from draft to promoted, built on record contracts and a quality bar every example must clear before import. The curation gates fail closed: legal-safety review, privacy review, and automatic skips for records contaminated by tool failures. And before training ever starts, evaluation holdouts are split from the training data — the examples used to measure the model are never the examples used to teach it, so the measurement stays honest.

The fine-tune itself

Here's the technical claim, stated plainly: fine-tuning in the Training Center is parameter-efficient, LoRA-class supervised fine-tuning of local open-source models, run on the firm's own hardware. Supervised fine-tuning means the model learns from curated example pairs — this input, this approved output — the most established and predictable way to specialize a model. It is not reinforcement learning, and we don't dress it up as something more exotic. The dataset and the resulting weights never leave the deployment. What comes out is the firm's asset, full stop.

Measurement you can inspect

After a training run, analytics, run traces, and evaluation against the held-out set show whether the model actually got better at the firm's work — not just different. That's the standard a firm should hold any model change to, including its own: show me the evaluation, on examples the model never saw, before this goes anywhere near a matter.

Notice what the pipeline implies about when to start. A firm doesn't need to be ready to train anything to begin walking this path — the feedback layer runs from day one, quietly accumulating the record of where the AI helped and where it didn't. Curation can start whenever the firm decides the record is worth mining. The point is that the option compounds: every corrected draft and every thumbs-down is signal a future dataset can draw on, sitting in your deployment, costing nothing until you act on it.

What a firm model is not

This is where we'd rather under-promise. Three things a firm-tuned model is not, from the people building the tooling:

• It is not a frontier-model killer. A model tuned on your practice gets sharper at your practice — your documents, your style, your recurring work. It does not outperform frontier models at general legal reasoning, and we won't tell you otherwise. That's why the spectrum exists: many firms will run frontier models for the hardest reasoning and their own model where its specialization pays.
• It is not fast or magic. Building a dataset worth training on takes real curation by real lawyers, and we make no promises about timelines. A small, clean, lawyer-reviewed dataset beats a large scraped one — and clean takes time.
• It is not a shortcut around judgment. Tuning shapes how the model writes and what patterns it knows. It does not make the model right, which is exactly why the governance layer doesn't move an inch for it.

Governance doesn't bend for the house model

The most important design decision in the Training Center may be this one: a firm-tuned model is still just a model under EdgeLex governance. The default-deny model policy applies to it. Capability binding applies to it. Evidence gates apply to it — its claims need typed, verified support or they're blocked, exactly like a frontier model's. Approval gates apply to it — it proposes, humans approve. The mutation-safety router evaluates it like any other model and will keep it away from destructive actions it can't carry.

The training records themselves reinforce this: examples enforce Lex's grounded, cite-your-sources, approval-gated persona, so a firm-tuned model gets sharper at your practice without getting looser about the rules. There is no trusted-insider path where "our model" earns a pass on evidence because the firm raised it. Familiarity is not verification — a principle law firms already apply to people, extended to models.

One more thing worth stating because clients ask: none of this happens by default. EdgeLex never trains on your data unless your firm deliberately curates a dataset and runs its own fine-tune — and the model that produces belongs to you, in your deployment, whether that deployment is EdgeLex-managed cloud or self-hosted on your own infrastructure. The Training Center is included with EdgeLex; firm-model training runs on the firm's infrastructure. The spectrum is there when you want it. Where you stand on it is, and stays, your firm's decision.